| name | openchoreo-developer |
| description | Application-level OpenChoreo work via the control-plane MCP server — deploying services, configuring workloads, consuming managed-infrastructure Resources, promoting releases, managing secret references, inspecting runtime. Use when the user says 'deploy my service', 'add a component', 'rebuild from source', 'use a database', 'promote to staging', 'rollback', or 'why is my pod crashing'. |
| metadata | {"version":"1.1.5"} |
OpenChoreo Developer Guide
Help an application developer ship and operate a service on OpenChoreo through the control-plane MCP server. Discover the live platform shape via MCP; read detailed references only when the task needs them.
Step 0 — Confirm MCP connectivity
Run list_namespaces. If the tool isn't reachable, tell the user the openchoreo-cp MCP server needs configuring per https://openchoreo.dev/docs/ai/mcp-servers/ and stop.
Step 1 — Load concepts (MANDATORY)
Read ./references/concepts.md in full before anything else. Not optional, even if the task looks simple — you'll get these wrong from memory. Load once per session; if you catch yourself acting without it, stop and load now.
Then load the matching reference for the task:
Step 2 — Load the PE skill for platform tasks
If the request touches ComponentType / ResourceType / Trait / Workflow (or cluster variants) or anything under What this skill cannot do, also load ../openchoreo-platform-engineer/SKILL.md. If the PE skill isn't installed, escalate.
What this skill can do
- Projects — create, update.
- Components — create, update,
patch_component for auto_deploy / parameters / traits / workflow / metadata.
- Workloads — image, ports, endpoints, env vars, files →
configure-workload.md
- Attach Traits — pick from the platform's catalog →
configure-workload.md
- Connect components — endpoint dependencies; platform injects env vars →
connect-components.md
- Consume Resources — managed-infrastructure dependencies (databases, queues, caches); platform injects outputs as env vars and file mounts via
dependencies.resources[] → use-a-resource.md
- SecretReferences — CRUD +
secretKeyRef consumption → manage-secrets.md. The ClusterSecretStore is PE-owned.
- Deploy and promote — bind a
ComponentRelease to an Environment, promote along the DeploymentPipeline → deploy-and-promote.md
- Per-environment overrides — replicas, resources, env vars, trait config on the ReleaseBinding →
override-per-environment.md
- Soft-undeploy / rollback —
update_release_binding release_state: Undeploy, or rebind to a prior ComponentRelease.
- Hard-delete developer resources —
delete_component, delete_workload, delete_release_binding, delete_project, delete_component_release. Destructive; confirm first. No delete_namespace — PE-side.
- Inspect runtime — Component / ReleaseBinding
status.conditions[] and status.endpoints[]; get_resource_tree to map a binding to its rendered K8s resources; get_resource_events / get_resource_logs for pod-level evidence; WorkflowRun logs and events → inspect-and-debug.md
- Discover platform resources (read-only) — ComponentTypes, ResourceTypes, Traits, Workflows, Environments, DeploymentPipelines, planes.
What this skill cannot do
Platform-side work: authoring ComponentTypes / ResourceTypes / Traits / Workflows, Environments, DeploymentPipelines, planes, authorization, gateway / secret-store / IdP config, observability setup, longer-horizon log / metric / trace queries (pod-level events and current logs are covered via get_resource_events / get_resource_logs).
Load ../openchoreo-platform-engineer/SKILL.md if it's installed; otherwise tell the user to escalate to a platform engineer. PE scope catalog: https://openchoreo.dev/docs/platform-engineer-guide/.
Tool surface
One MCP server: openchoreo-cp. Throughout this skill, tools are referenced by bare name (e.g. get_component); your agent wraps with its prefix (Claude Code uses mcp__openchoreo-cp__<tool>).
ComponentType / ResourceType / Trait / Workflow and the plane tools are scope-collapsed: one tool with a scope arg — "namespace" (default) or "cluster" for the platform-wide Cluster* resource. This skill always uses the canonical name + scope. The old *_cluster_* names still exist as deprecated aliases (banner in v1.1, hidden in v1.2, removed in v1.3) and can be used alternatively against a v1.1 server — but prefer the canonical form.
The resource toolset (dev-facing, enabled by default) covers Resource CRUD plus scope-collapsed reads on (Cluster)ResourceType. ResourceReleaseBinding CRUD sits in the deployment toolset alongside ReleaseBinding. The resource_release_binding create/update calls are how a Resource gets deployed into an env (or promoted to a new release).
Working style
- Live cluster output beats memory. Discover via MCP first; don't guess available ComponentTypes / Traits / Workflows / Environments / field names.
- Schema-first authoring.
get_workload_schema, or get_component_type_schema / get_trait_schema with scope: "cluster" for platform-wide standards, before writing a spec from scratch.
Stable guardrails
- All work goes through the control-plane MCP server. If a task can't be done via MCP, it's platform-side — hand off.
- Third-party / public apps: default to BYO image. Source builds commonly fail on third-party Dockerfiles using
ARG BUILDPLATFORM (exit 125). Switch to BYO immediately if you see it.
- Before deploying any third-party app: fetch the official manifests and extract every required env var — dependencies inject service addresses but not
PORT, feature flags, or vendor SDK disable flags.
- A handed-over migration plan is the spec. When the user supplies a migration/onboarding plan, take namespace, env var placement (static Workload env vs per-env
workloadOverrides), and wiring decisions from it. Deviate only out loud — state what you're changing and why before acting, never silently substitute.
- Author types to fit the app — never re-shape the app to fit shipped types. When the plan (or the app) needs a
ComponentType / ResourceType / Trait that isn't installed or that ships but doesn't meet the requirement, the fix is to author it — load ../openchoreo-platform-engineer/SKILL.md and create a fit-for-purpose type. Do not work around the gap by: re-modeling a managed dependency as a plain Component because the shipped ResourceType doesn't fit (it's still a Resource — author a ResourceType that reproduces what the dependency needs); dropping a capability the app needs because no shipped Trait provides it (author the Trait); or forcing the workload onto a shipped ComponentType whose shape or defaults don't fit (author the ComponentType). Reuse a shipped type only when it genuinely meets the contract. The needs drive the types, not the reverse.
- Set resources from the source, not the ComponentType's defaults. Use the workload's source-declared requests/limits (CPU and memory). If the source declares none, pick a plausible default for the workload class — never inherit a generic CT's defaults, which are typically demo-grade and under-provision. When you author the CT, bake sensible defaults into its
environmentConfigs; otherwise set them via the ReleaseBinding's per-env config.
- Injected connection values are literal — there is no
$(VAR) interpolation. Dependency envBindings inject each connection value (address / host / port / basePath) as the literal env var you name; the platform does not expand shell/Kubernetes $(VAR) references inside other env values, so value: "http://$(HOST):$(PORT)" reaches the container verbatim. Bind the exact env var the app reads. If the app needs a composed string the injected vars don't directly provide (a full URI, a bare host:port with no scheme, a scheme prefix), set that composed value as a literal env var — typically a per-env workloadOverride using the resolved address — rather than referencing $(VAR).
- A missing tool means version skew, not absence. When a documented MCP tool isn't found, check the server version against the cluster before concluding the surface doesn't exist — report the mismatch to the user, then fall back or escalate.
Anti-patterns
- Skipping the recipe. Before any new operation (new CRD kind this turn, lifecycle action, runtime inspection) — re-scan the recipe index above, load the matching recipe (one Read call), THEN call MCP / kubectl. Skipping is how kubectl falls multiply and existing MCP tools get missed. Concept references aren't enough — recipes name the right tool calls in sequence.
- Running every discovery call before checking the resource already implicated.
- Writing specs from memory when
get_*_schema / get_* can reveal the current shape.
- Guessing deployed URLs instead of reading
ReleaseBinding.status.endpoints[].
- Treating a platform-side failure as an app-only problem after the evidence points elsewhere.
- Creating source-build components for third-party apps that have pre-built images.
- Setting
visibility: external on a service-to-service dependency in the same project — project is the default.
- Treating
Ready=True as "working". Ready means reconciled, not functional. A crash-looping container can flap Ready, and a stable container can be Ready while misconfigured (env vars bound to wrong names, deps resolving to nowhere). Confirm with get_resource_tree → get_resource_events / get_resource_logs and an actual endpoint hit.